R1-2500644
discussion
Specification Support for AI/ML CSI prediction
From Sony
Summary
Sony presents a technical contribution for 3GPP RAN1 regarding specification support for AI/ML-based CSI prediction in the NR Air Interface. The document identifies three key areas requiring standardization: the temporal framework for prediction slots, the specific form of channel data inputs/outputs, and the configuration of measurement resources for monitoring. It contains 1 observation and 3 proposals aimed at defining how UEs should predict CSI for future time slots and how network monitoring metrics should be applied.
Position
Sony proposes that RAN1 specify a CSI prediction framework allowing the UE to carry out predictions for [N] prediction time slots simultaneously, addressing the uncertainty of gNB scheduler decisions regarding k0 and SLIV. They present two options for slot selection: gNB-configured candidate slots or UE-independent selection within an implied maximum time. Regarding model inputs, Sony proposes specifying either CSI-RS values with NMSE monitoring or channel matrices with NMSE/SGCS monitoring, emphasizing that ground truth measurements must be available in the prediction slot. They further propose that RAN1 specify the configuration of measurement resources for both model inference inputs and the related ground truths required for monitoring. Sony argues that model training inputs may need to include the prediction slot itself to ensure accurate temporal alignment.
Key proposals
- Proposal 1 (Sec 3): RAN1 should specify a CSI prediction framework that allows CSI prediction at the UE to be carried out for [N] prediction time slots at a time.
- Proposal 2 (Sec 4): RAN1 should specify the form of channel data for data collection, model training, model input at inference and monitoring of CSI prediction model and the consequent model monitoring metric.
- Proposal 2 Option A (Sec 4): Use CSI-RS value with NMSE for monitoring metric.
- Proposal 2 Option B (Sec 4): Use Channel matrix derived from CSI-RS measurements with NMSE or SGCS for monitoring metric.
- Proposal 3 (Sec 5): RAN1 should specify how measurement resources for both model inference input and related ground truths for monitoring are configured.
- Observation 1 (Sec 3): Model training for CSI prediction may have to include amongst its inputs the prediction slot.
- Option 1 (Sec 3): gNB configures multiple prediction slots for the UE to predict CSI.
- Option 2 (Sec 3): UE independently decides multiple prediction slots, potentially uniformly spaced within a configured or implied maximum time interval.
- Option A Detail (Sec 4): Model input is received complex values of CSI-RS; output is predicted CSI-RS value; monitoring compares predicted vs measured CSI-RS using NMSE.
- Option B Detail (Sec 4): Model input is a channel matrix generated from CSI-RS measurements; output is a predicted channel matrix; monitoring uses NMSE or SGCS.
- Resource Config Context (Sec 5): Ground truth measurement resources are needed in the prediction slot to compare against predicted CSI values for accuracy assessment.
- Scheduling Context (Sec 3): The framework must account for the fact that the UE does not know k0 (time domain resource allocation) at prediction time, necessitating prediction for multiple candidate slots.